CrediLume 2.0

project architecture

Inspiration

It was inspired by a simple but urgent problem: most people make major borrowing decisions without clear, reliable guidance. We saw that borrowers, especially in semi-urban and rural contexts, often get numbers but not understanding. Our goal was to build something that makes lending decisions transparent, practical, and understandable before a user commits to a loan.

What it does

It is an AI-assisted loan intelligence platform that helps users estimate EMI, assess affordability, and understand likely approval outcomes with clear explanations. Instead of only returning a score, it provides context, decision reasoning, and actionable recommendations. It also includes a dedicated agriculture-focused flow so farmer borrowers can be evaluated with logic that reflects their real repayment patterns and income cycles.

How we built it

We built Credilume 2.0 using a Flask-based backend connected to a frontend designed for fast, clear interaction. The core combines deterministic financial calculations with model-driven risk estimation, and then applies a policy decision layer to produce outputs such as approve, manual review, or decline. We also integrated advisory capabilities with fallback behavior so the platform remains stable and useful even if external AI providers are unavailable.

Challenges we ran into

One of the biggest challenges was balancing predictive intelligence with explainability, because users need both confidence and clarity. We also had to carefully handle edge cases in financial inputs and ensure that policy logic stayed consistent across user flows. Designing for both standard consumer lending and farmer-specific credit journeys added complexity, since each requires different assumptions and interpretation of risk.

Accomplishments that we're proud of

We are proud that Ledger goes beyond a black-box prediction and delivers transparent, decision-ready insights. The agriculture mode is another major achievement because it is tailored rather than generic, making the product more useful for real-world lending diversity. We are also proud of delivering a complete, deployable platform with unified branding, polished UX, and a strong backend foundation.

What we learned

We learned that trust in fintech tools is built as much through communication as through model quality. Users engage better when outputs are concrete, interpretable, and tied to actions they can take next. We also learned the importance of hybrid architecture: deterministic logic, ML predictions, and plain-language recommendations together create a stronger and more reliable experience than any single layer alone.

What's next for CrediLume 2.0

Next, we plan to deepen underwriting quality through richer data integrations, improve personalization through progress-based eligibility roadmaps, and expand accessibility with multilingual support. We also want to build lender-facing analytics for portfolio insight, strengthen governance with audit and fairness tooling, and continue evolving agri-credit intelligence using seasonal and regional signals for more context-aware lending decisions.

Presentation Link: https://docs.google.com/presentation/d/1TNQnnqk9_p3AwKGijLaBPgWapZPzNSn9/edit?usp=sharing&ouid=117918945453140069608&rtpof=true&sd=true